Transformer transfer learning emotion detection model: synchronizing socially agreed and self-reported emotions in big data

Abstract

Tactics to determine the emotions of authors of texts such as Twitter messages often rely on multiple annotators who label relatively small data sets of text passages. An alternative method gathers large text databases that contain the authors’ self-reported emotions, to which artificial intelligence, machine learning, and natural language processing tools can be applied. Both approaches have strength and weaknesses. Emotions evaluated by a few human annotators are susceptible to idiosyncratic biases that reflect the characteristics of the annotators. But models based on large, self-reported emotion data sets may overlook subtle, social emotions that human annotators can recognize. In seeking to establish a means to train emotion detection models so that they can achieve good performance in different contexts, the current study proposes a novel transformer transfer learning approach that parallels human development stages: (1) detect emotions reported by the texts’ authors and (2) synchronize the model with social emotions identified in annotator-rated emotion data sets. The analysis, based on a large, novel, self-reported emotion data set (n = 3,654,544) and applied to 10 previously published data sets, shows that the transfer learning emotion model achieves relatively strong performance.